Sequoia's Abhishek Malani on the AI poker-bot tournament: 7,000 applicants, 700 competitors, and a Sunday final

Abhishek Malani joined Sequoia Capital roughly six weeks ago with an idea born on a weekend: could you build a poker bot quickly with modern AI models? The answer, it turns out, is yes — and the experiment grew fast enough to become a full tournament, with a final this Sunday.

The numbers are striking. The competition drew 7,000 applicants, whittled down to 700 active competitors, with the top 20 teams from Boston, the Bay Area, and Pittsburgh advancing to the finals. Sequoia built it in partnership with OpenAI and Citadel Securities, the latter brought in because its core business involves strategy under hidden information. The first call Malani made was to Noam Brown, the researcher who pioneered poker-solving AI at CMU starting in 2017.

“About a month and a half ago I was playing a lot of poker with friends, realizing these new models are coming out, and holy crap, they can do a lot of things. So come into partner meeting at Sequoia one day, telling people this is something interesting — and immediately Sean, Alfred, a lot of the partners are like, we need to do something about this. Over now 7,000 people have applied. We have over 700 competing. Finals is gonna have your top 20 teams from Boston, The Bay, Pittsburgh competing.”

How competitors are approaching it

The field ranges widely. At the low end, some entrants simply ask Claude or Codex to generate a bot and hope for the best. The more serious competitors are building classification models that read opponent behavior within roughly 15 hands — identifying whether a rival bot is loose or tight — and then adjusting strategy in real time to exploit the pattern. Others are building neural networks from scratch. The CMU regional winner used exactly this classification approach.

A one-millisecond response-time constraint rules out brute-force test-time compute as a shortcut. That cap keeps the competition grounded in architecture and strategy rather than raw spend.

Custom variants as the real differentiator

Each regional uses a different poker variant, designed specifically to prevent teams from downloading an existing solver and wrapping it. At the Boston regional, for example, players started with the standard two hole cards but faced a five-card flop — then had to bid against the other bots at the table for which three of those five cards would actually become the flop. That bidding mechanic changes the entire decision tree in ways that a standard poker engine cannot handle. Malani frames this as the core design principle: novel enough that pre-existing tools break.

The talent angle

For Sequoia, the tournament is a recruiting funnel as much as a technical exercise. Malani describes competitors building Tesla coils, designing custom PCBs, and writing proprietary trading software while still in high school. Whether the best performers end up founding companies, joining OpenAI for research, or trading at Citadel is secondary — the goal is early access to high-agency engineers before anyone else has mapped them.

Malani built his own bot on a weekend to prove the concept was feasible. He is not competing in the finals.